The core claims sit in B25J9 and B25J19, the manipulator and robot-safety subclasses, alongside F16P3 for mechanical safety devices. In this dataset every one of the 71 matched families falls under B25J, with control-system logic (G05B) layered on top in a smaller share of filings. AI-based control under G06N and standalone force or torque sensing under G01L and G01D appear only occasionally, which marks them as the less-crowded adjacent classes rather than the primary battleground.

Filing activity in this space is spread across Chinese, Korean, US and European applicants rather than concentrated in one dominant filer. Receiving-office data shows China and the European Patent Office each handling 18 records and the US handling 17, so no single jurisdiction — and by extension no single home market — clearly leads. The most-cited individual records include both a US assignee and several Chinese assignees, reinforcing that this is a multi-region contest rather than a one-company field.

Not by the numbers so far: filings climbed from 4 in 2017 to a peak of 11 in 2022, and have not gone above that since. That looks like a plateau rather than continued growth, though the most recent one to two years are understated because publication typically lags filing by around 18 months. Treat the 2022 peak as the clearest read on real momentum and expect the newest filings to be revised upward as they publish.

US12491636B2, granted to NEUROMEKA on 2025-12-09, claims a collaborative robot architecture combining a main-body robot with an additional-shaft robot that moves it, plus a processor unit that calculates an external-force value from both robots' signals and compares it against a predetermined collision-detection boundary value to determine a collision. It is specific to that additional-shaft configuration and that boundary-value comparison method, so it does not cover collision detection in general — single-axis arms or speed-separation-based safety monitoring sit outside its literal scope. Anyone designing a multi-axis cobot with an auxiliary travel axis should check this claim closely.

The thinnest claimed territory sits in AI-based collision prediction (G06N, 3 records), standalone force/pressure measurement (G01L, 1 record) and non-electric variable control (G05D, 1 record), all dwarfed by the 71 records under the core B25J mechanical-safety class. That gap suggests predictive or learned collision-avoidance models, as opposed to fixed boundary-value comparisons, remain comparatively open. It is a filing-density signal, not proof the approach is technically superior — thin classes can also mean the approach is harder to productise.